[refactor] CogVideoX followups + tiled decoding support (#9150)

* refactor context parallel cache; update torch compile time benchmark

* add tiling support

* make style

* remove num_frames % 8 == 0 requirement

* update default num_frames to original value

* add explanations + refactor

* update torch compile example

* update docs

* update

* clean up if-statements

* address review comments

* add test for vae tiling

* update docs

* update docs

* update docstrings

* add modeling test for cogvideox transformer

* make style

docs/source/en/api/pipelines/cogvideox.md

@@ -15,9 +15,7 @@
 
 # CogVideoX
 
-<!-- TODO: update paper with ArXiv link when ready. -->
-
-[CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer](https://github.com/THUDM/CogVideo/blob/main/resources/CogVideoX.pdf) from Tsinghua University & ZhipuAI.
+[CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer](https://arxiv.org/abs/2408.06072) from Tsinghua University & ZhipuAI, by Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, Jie Tang.
 
 The abstract from the paper is:
 
@@ -43,43 +41,42 @@ from diffusers import CogVideoXPipeline
 from diffusers.utils import export_to_video
 
 pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-2b").to("cuda")
-prompt = (
-    "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. "
-    "The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other "
-    "pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, "
-    "casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. "
-    "The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical "
-    "atmosphere of this unique musical performance."
-)
-video = pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
-export_to_video(video, "output.mp4", fps=8)
 ```
 
-Then change the memory layout of the pipelines `transformer` and `vae` components to `torch.channels-last`:
+Then change the memory layout of the pipelines `transformer` component to `torch.channels_last`:
 
 ```python
-pipeline.transformer.to(memory_format=torch.channels_last)
-pipeline.vae.to(memory_format=torch.channels_last)
+pipe.transformer.to(memory_format=torch.channels_last)
 ```
 
 Finally, compile the components and run inference:
 
 ```python
-pipeline.transformer = torch.compile(pipeline.transformer)
-pipeline.vae.decode = torch.compile(pipeline.vae.decode)
+pipe.transformer = torch.compile(pipeline.transformer, mode="max-autotune", fullgraph=True)
 
-# CogVideoX works very well with long and well-described prompts
+# CogVideoX works well with long and well-described prompts
 prompt = "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical atmosphere of this unique musical performance."
-video = pipeline(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
+video = pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
 ```
 
-The [benchmark](TODO: link) results on an 80GB A100 machine are:
+The [benchmark](https://gist.github.com/a-r-r-o-w/5183d75e452a368fd17448fcc810bd3f) results on an 80GB A100 machine are:
 
 ```
-Without torch.compile(): Average inference time: TODO seconds.
-With torch.compile(): Average inference time: TODO seconds.
+Without torch.compile(): Average inference time: 96.89 seconds.
+With torch.compile(): Average inference time: 76.27 seconds.
 ```
 
+### Memory optimization
+
+CogVideoX requires about 19 GB of GPU memory to decode 49 frames (6 seconds of video at 8 FPS) with output resolution 720x480 (W x H), which makes it not possible to run on consumer GPUs or free-tier T4 Colab. The following memory optimizations could be used to reduce the memory footprint. For replication, you can refer to [this](https://gist.github.com/a-r-r-o-w/3959a03f15be5c9bd1fe545b09dfcc93) script.
+
+- `pipe.enable_model_cpu_offload()`:
+  - Without enabling cpu offloading, memory usage is `33 GB`
+  - With enabling cpu offloading, memory usage is `19 GB`
+- `pipe.vae.enable_tiling()`:
+  - With enabling cpu offloading and tiling, memory usage is `11 GB`
+- `pipe.vae.enable_slicing()`
+
 ## CogVideoXPipeline
 
 [[autodoc]] CogVideoXPipeline

src/diffusers/models/autoencoders/autoencoder_kl_cogvideox.py

@@ -36,7 +36,7 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
 class CogVideoXSafeConv3d(nn.Conv3d):
-    """
+    r"""
     A 3D convolution layer that splits the input tensor into smaller parts to avoid OOM in CogVideoX Model.
     """
 
@@ -68,12 +68,12 @@ class CogVideoXCausalConv3d(nn.Module):
     r"""A 3D causal convolution layer that pads the input tensor to ensure causality in CogVideoX Model.
 
     Args:
-        in_channels (int): Number of channels in the input tensor.
-        out_channels (int): Number of output channels.
-        kernel_size (Union[int, Tuple[int, int, int]]): Size of the convolutional kernel.
-        stride (int, optional): Stride of the convolution. Default is 1.
-        dilation (int, optional): Dilation rate of the convolution. Default is 1.
-        pad_mode (str, optional): Padding mode. Default is "constant".
+        in_channels (`int`): Number of channels in the input tensor.
+        out_channels (`int`): Number of output channels produced by the convolution.
+        kernel_size (`int` or `Tuple[int, int, int]`): Kernel size of the convolutional kernel.
+        stride (`int`, defaults to `1`): Stride of the convolution.
+        dilation (`int`, defaults to `1`): Dilation rate of the convolution.
+        pad_mode (`str`, defaults to `"constant"`): Padding mode.
     """
 
     def __init__(
@@ -118,19 +118,12 @@ class CogVideoXCausalConv3d(nn.Module):
         self.conv_cache = None
 
     def fake_context_parallel_forward(self, inputs: torch.Tensor) -> torch.Tensor:
-        dim = self.temporal_dim
         kernel_size = self.time_kernel_size
-        if kernel_size == 1:
-            return inputs
-
-        inputs = inputs.transpose(0, dim)
-
-        if self.conv_cache is not None:
-            inputs = torch.cat([self.conv_cache.transpose(0, dim).to(inputs.device), inputs], dim=0)
-        else:
-            inputs = torch.cat([inputs[:1]] * (kernel_size - 1) + [inputs], dim=0)
-
-        inputs = inputs.transpose(0, dim).contiguous()
+        if kernel_size > 1:
+            cached_inputs = (
+                [self.conv_cache] if self.conv_cache is not None else [inputs[:, :, :1]] * (kernel_size - 1)
+            )
+            inputs = torch.cat(cached_inputs + [inputs], dim=2)
         return inputs
 
     def _clear_fake_context_parallel_cache(self):
@@ -138,16 +131,17 @@ class CogVideoXCausalConv3d(nn.Module):
         self.conv_cache = None
 
     def forward(self, inputs: torch.Tensor) -> torch.Tensor:
-        input_parallel = self.fake_context_parallel_forward(inputs)
+        inputs = self.fake_context_parallel_forward(inputs)
 
         self._clear_fake_context_parallel_cache()
-        self.conv_cache = input_parallel[:, :, -self.time_kernel_size + 1 :].contiguous().detach().clone().cpu()
+        # Note: we could move these to the cpu for a lower maximum memory usage but its only a few
+        # hundred megabytes and so let's not do it for now
+        self.conv_cache = inputs[:, :, -self.time_kernel_size + 1 :].clone()
 
         padding_2d = (self.width_pad, self.width_pad, self.height_pad, self.height_pad)
-        input_parallel = F.pad(input_parallel, padding_2d, mode="constant", value=0)
+        inputs = F.pad(inputs, padding_2d, mode="constant", value=0)
 
-        output_parallel = self.conv(input_parallel)
-        output = output_parallel
+        output = self.conv(inputs)
         return output
 
 
@@ -163,6 +157,8 @@ class CogVideoXSpatialNorm3D(nn.Module):
             The number of channels for input to group normalization layer, and output of the spatial norm layer.
         zq_channels (`int`):
             The number of channels for the quantized vector as described in the paper.
+        groups (`int`):
+            Number of groups to separate the channels into for group normalization.
     """
 
     def __init__(
@@ -197,17 +193,26 @@ class CogVideoXResnetBlock3D(nn.Module):
     A 3D ResNet block used in the CogVideoX model.
 
     Args:
-        in_channels (int): Number of input channels.
-        out_channels (Optional[int], optional):
-            Number of output channels. If None, defaults to `in_channels`. Default is None.
-        dropout (float, optional): Dropout rate. Default is 0.0.
-        temb_channels (int, optional): Number of time embedding channels. Default is 512.
-        groups (int, optional): Number of groups for group normalization. Default is 32.
-        eps (float, optional): Epsilon value for normalization layers. Default is 1e-6.
-        non_linearity (str, optional): Activation function to use. Default is "swish".
-        conv_shortcut (bool, optional): If True, use a convolutional shortcut. Default is False.
-        spatial_norm_dim (Optional[int], optional): Dimension of the spatial normalization. Default is None.
-        pad_mode (str, optional): Padding mode. Default is "first".
+        in_channels (`int`):
+            Number of input channels.
+        out_channels (`int`, *optional*):
+            Number of output channels. If None, defaults to `in_channels`.
+        dropout (`float`, defaults to `0.0`):
+            Dropout rate.
+        temb_channels (`int`, defaults to `512`):
+            Number of time embedding channels.
+        groups (`int`, defaults to `32`):
+            Number of groups to separate the channels into for group normalization.
+        eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        non_linearity (`str`, defaults to `"swish"`):
+            Activation function to use.
+        conv_shortcut (bool, defaults to `False`):
+            Whether or not to use a convolution shortcut.
+        spatial_norm_dim (`int`, *optional*):
+            The dimension to use for spatial norm if it is to be used instead of group norm.
+        pad_mode (str, defaults to `"first"`):
+            Padding mode.
     """
 
     def __init__(
@@ -309,18 +314,28 @@ class CogVideoXDownBlock3D(nn.Module):
     A downsampling block used in the CogVideoX model.
 
     Args:
-        in_channels (int): Number of input channels.
-        out_channels (int): Number of output channels.
-        temb_channels (int): Number of time embedding channels.
-        dropout (float, optional): Dropout rate. Default is 0.0.
-        num_layers (int, optional): Number of layers in the block. Default is 1.
-        resnet_eps (float, optional): Epsilon value for the ResNet layers. Default is 1e-6.
-        resnet_act_fn (str, optional): Activation function for the ResNet layers. Default is "swish".
-        resnet_groups (int, optional): Number of groups for group normalization in the ResNet layers. Default is 32.
-        add_downsample (bool, optional): If True, add a downsampling layer at the end of the block. Default is True.
-        downsample_padding (int, optional): Padding for the downsampling layer. Default is 0.
-        compress_time (bool, optional): If True, apply temporal compression. Default is False.
-        pad_mode (str, optional): Padding mode. Default is "first".
+        in_channels (`int`):
+            Number of input channels.
+        out_channels (`int`, *optional*):
+            Number of output channels. If None, defaults to `in_channels`.
+        temb_channels (`int`, defaults to `512`):
+            Number of time embedding channels.
+        num_layers (`int`, defaults to `1`):
+            Number of resnet layers.
+        dropout (`float`, defaults to `0.0`):
+            Dropout rate.
+        resnet_eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        resnet_act_fn (`str`, defaults to `"swish"`):
+            Activation function to use.
+        resnet_groups (`int`, defaults to `32`):
+            Number of groups to separate the channels into for group normalization.
+        add_downsample (`bool`, defaults to `True`):
+            Whether or not to use a downsampling layer. If not used, output dimension would be same as input dimension.
+        compress_time (`bool`, defaults to `False`):
+            Whether or not to downsample across temporal dimension.
+        pad_mode (str, defaults to `"first"`):
+            Padding mode.
     """
 
     _supports_gradient_checkpointing = True
@@ -405,15 +420,24 @@ class CogVideoXMidBlock3D(nn.Module):
     A middle block used in the CogVideoX model.
 
     Args:
-        in_channels (int): Number of input channels.
-        temb_channels (int): Number of time embedding channels.
-        dropout (float, optional): Dropout rate. Default is 0.0.
-        num_layers (int, optional): Number of layers in the block. Default is 1.
-        resnet_eps (float, optional): Epsilon value for the ResNet layers. Default is 1e-6.
-        resnet_act_fn (str, optional): Activation function for the ResNet layers. Default is "swish".
-        resnet_groups (int, optional): Number of groups for group normalization in the ResNet layers. Default is 32.
-        spatial_norm_dim (Optional[int], optional): Dimension of the spatial normalization. Default is None.
-        pad_mode (str, optional): Padding mode. Default is "first".
+        in_channels (`int`):
+            Number of input channels.
+        temb_channels (`int`, defaults to `512`):
+            Number of time embedding channels.
+        dropout (`float`, defaults to `0.0`):
+            Dropout rate.
+        num_layers (`int`, defaults to `1`):
+            Number of resnet layers.
+        resnet_eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        resnet_act_fn (`str`, defaults to `"swish"`):
+            Activation function to use.
+        resnet_groups (`int`, defaults to `32`):
+            Number of groups to separate the channels into for group normalization.
+        spatial_norm_dim (`int`, *optional*):
+            The dimension to use for spatial norm if it is to be used instead of group norm.
+        pad_mode (str, defaults to `"first"`):
+            Padding mode.
     """
 
     _supports_gradient_checkpointing = True
@@ -480,19 +504,30 @@ class CogVideoXUpBlock3D(nn.Module):
     An upsampling block used in the CogVideoX model.
 
     Args:
-        in_channels (int): Number of input channels.
-        out_channels (int): Number of output channels.
-        temb_channels (int): Number of time embedding channels.
-        dropout (float, optional): Dropout rate. Default is 0.0.
-        num_layers (int, optional): Number of layers in the block. Default is 1.
-        resnet_eps (float, optional): Epsilon value for the ResNet layers. Default is 1e-6.
-        resnet_act_fn (str, optional): Activation function for the ResNet layers. Default is "swish".
-        resnet_groups (int, optional): Number of groups for group normalization in the ResNet layers. Default is 32.
-        spatial_norm_dim (int, optional): Dimension of the spatial normalization. Default is 16.
-        add_upsample (bool, optional): If True, add an upsampling layer at the end of the block. Default is True.
-        upsample_padding (int, optional): Padding for the upsampling layer. Default is 1.
-        compress_time (bool, optional): If True, apply temporal compression. Default is False.
-        pad_mode (str, optional): Padding mode. Default is "first".
+        in_channels (`int`):
+            Number of input channels.
+        out_channels (`int`, *optional*):
+            Number of output channels. If None, defaults to `in_channels`.
+        temb_channels (`int`, defaults to `512`):
+            Number of time embedding channels.
+        dropout (`float`, defaults to `0.0`):
+            Dropout rate.
+        num_layers (`int`, defaults to `1`):
+            Number of resnet layers.
+        resnet_eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+        resnet_act_fn (`str`, defaults to `"swish"`):
+            Activation function to use.
+        resnet_groups (`int`, defaults to `32`):
+            Number of groups to separate the channels into for group normalization.
+        spatial_norm_dim (`int`, defaults to `16`):
+            The dimension to use for spatial norm if it is to be used instead of group norm.
+        add_upsample (`bool`, defaults to `True`):
+            Whether or not to use a upsampling layer. If not used, output dimension would be same as input dimension.
+        compress_time (`bool`, defaults to `False`):
+            Whether or not to downsample across temporal dimension.
+        pad_mode (str, defaults to `"first"`):
+            Padding mode.
     """
 
     def __init__(
@@ -587,14 +622,12 @@ class CogVideoXEncoder3D(nn.Module):
             options.
         block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
             The number of output channels for each block.
+        act_fn (`str`, *optional*, defaults to `"silu"`):
+            The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
         layers_per_block (`int`, *optional*, defaults to 2):
             The number of layers per block.
         norm_num_groups (`int`, *optional*, defaults to 32):
             The number of groups for normalization.
-        act_fn (`str`, *optional*, defaults to `"silu"`):
-            The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
-        double_z (`bool`, *optional*, defaults to `True`):
-            Whether to double the number of output channels for the last block.
     """
 
     _supports_gradient_checkpointing = True
@@ -723,14 +756,12 @@ class CogVideoXDecoder3D(nn.Module):
             The types of up blocks to use. See `~diffusers.models.unet_2d_blocks.get_up_block` for available options.
         block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
             The number of output channels for each block.
+        act_fn (`str`, *optional*, defaults to `"silu"`):
+            The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
         layers_per_block (`int`, *optional*, defaults to 2):
             The number of layers per block.
         norm_num_groups (`int`, *optional*, defaults to 32):
             The number of groups for normalization.
-        act_fn (`str`, *optional*, defaults to `"silu"`):
-            The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
-        norm_type (`str`, *optional*, defaults to `"group"`):
-            The normalization type to use. Can be either `"group"` or `"spatial"`.
     """
 
     _supports_gradient_checkpointing = True
@@ -911,7 +942,8 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         norm_eps: float = 1e-6,
         norm_num_groups: int = 32,
         temporal_compression_ratio: float = 4,
-        sample_size: int = 256,
+        sample_height: int = 480,
+        sample_width: int = 720,
         scaling_factor: float = 1.15258426,
         shift_factor: Optional[float] = None,
         latents_mean: Optional[Tuple[float]] = None,
@@ -950,25 +982,105 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
         self.use_slicing = False
         self.use_tiling = False
 
-        self.tile_sample_min_size = self.config.sample_size
-        sample_size = (
-            self.config.sample_size[0]
-            if isinstance(self.config.sample_size, (list, tuple))
-            else self.config.sample_size
+        # Can be increased to decode more latent frames at once, but comes at a reasonable memory cost and it is not
+        # recommended because the temporal parts of the VAE, here, are tricky to understand.
+        # If you decode X latent frames together, the number of output frames is:
+        #     (X + (2 conv cache) + (2 time upscale_1) + (4 time upscale_2) - (2 causal conv downscale)) => X + 6 frames
+        #
+        # Example with num_latent_frames_batch_size = 2:
+        #     - 12 latent frames: (0, 1), (2, 3), (4, 5), (6, 7), (8, 9), (10, 11) are processed together
+        #         => (12 // 2 frame slices) * ((2 num_latent_frames_batch_size) + (2 conv cache) + (2 time upscale_1) + (4 time upscale_2) - (2 causal conv downscale))
+        #         => 6 * 8 = 48 frames
+        #     - 13 latent frames: (0, 1, 2) (special case), (3, 4), (5, 6), (7, 8), (9, 10), (11, 12) are processed together
+        #         => (1 frame slice) * ((3 num_latent_frames_batch_size) + (2 conv cache) + (2 time upscale_1) + (4 time upscale_2) - (2 causal conv downscale)) +
+        #            ((13 - 3) // 2) * ((2 num_latent_frames_batch_size) + (2 conv cache) + (2 time upscale_1) + (4 time upscale_2) - (2 causal conv downscale))
+        #         => 1 * 9 + 5 * 8 = 49 frames
+        # It has been implemented this way so as to not have "magic values" in the code base that would be hard to explain. Note that
+        # setting it to anything other than 2 would give poor results because the VAE hasn't been trained to be adaptive with different
+        # number of temporal frames.
+        self.num_latent_frames_batch_size = 2
+
+        # We make the minimum height and width of sample for tiling half that of the generally supported
+        self.tile_sample_min_height = sample_height // 2
+        self.tile_sample_min_width = sample_width // 2
+        self.tile_latent_min_height = int(
+            self.tile_sample_min_height / (2 ** (len(self.config.block_out_channels) - 1))
         )
-        self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1)))
-        self.tile_overlap_factor = 0.25
+        self.tile_latent_min_width = int(self.tile_sample_min_width / (2 ** (len(self.config.block_out_channels) - 1)))
+
+        # These are experimental overlap factors that were chosen based on experimentation and seem to work best for
+        # 720x480 (WxH) resolution. The above resolution is the strongly recommended generation resolution in CogVideoX
+        # and so the tiling implementation has only been tested on those specific resolutions.
+        self.tile_overlap_factor_height = 1 / 6
+        self.tile_overlap_factor_width = 1 / 5
 
     def _set_gradient_checkpointing(self, module, value=False):
         if isinstance(module, (CogVideoXEncoder3D, CogVideoXDecoder3D)):
             module.gradient_checkpointing = value
 
-    def clear_fake_context_parallel_cache(self):
+    def _clear_fake_context_parallel_cache(self):
         for name, module in self.named_modules():
             if isinstance(module, CogVideoXCausalConv3d):
                 logger.debug(f"Clearing fake Context Parallel cache for layer: {name}")
                 module._clear_fake_context_parallel_cache()
 
+    def enable_tiling(
+        self,
+        tile_sample_min_height: Optional[int] = None,
+        tile_sample_min_width: Optional[int] = None,
+        tile_overlap_factor_height: Optional[float] = None,
+        tile_overlap_factor_width: Optional[float] = None,
+    ) -> None:
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+
+        Args:
+            tile_sample_min_height (`int`, *optional*):
+                The minimum height required for a sample to be separated into tiles across the height dimension.
+            tile_sample_min_width (`int`, *optional*):
+                The minimum width required for a sample to be separated into tiles across the width dimension.
+            tile_overlap_factor_height (`int`, *optional*):
+                The minimum amount of overlap between two consecutive vertical tiles. This is to ensure that there are
+                no tiling artifacts produced across the height dimension. Must be between 0 and 1. Setting a higher
+                value might cause more tiles to be processed leading to slow down of the decoding process.
+            tile_overlap_factor_width (`int`, *optional*):
+                The minimum amount of overlap between two consecutive horizontal tiles. This is to ensure that there
+                are no tiling artifacts produced across the width dimension. Must be between 0 and 1. Setting a higher
+                value might cause more tiles to be processed leading to slow down of the decoding process.
+        """
+        self.use_tiling = True
+        self.tile_sample_min_height = tile_sample_min_height or self.tile_sample_min_height
+        self.tile_sample_min_width = tile_sample_min_width or self.tile_sample_min_width
+        self.tile_latent_min_height = int(
+            self.tile_sample_min_height / (2 ** (len(self.config.block_out_channels) - 1))
+        )
+        self.tile_latent_min_width = int(self.tile_sample_min_width / (2 ** (len(self.config.block_out_channels) - 1)))
+        self.tile_overlap_factor_height = tile_overlap_factor_height or self.tile_overlap_factor_height
+        self.tile_overlap_factor_width = tile_overlap_factor_width or self.tile_overlap_factor_width
+
+    def disable_tiling(self) -> None:
+        r"""
+        Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.use_tiling = False
+
+    def enable_slicing(self) -> None:
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.use_slicing = True
+
+    def disable_slicing(self) -> None:
+        r"""
+        Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.use_slicing = False
+
     @apply_forward_hook
     def encode(
         self, x: torch.Tensor, return_dict: bool = True
@@ -993,8 +1105,34 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             return (posterior,)
         return AutoencoderKLOutput(latent_dist=posterior)
 
+    def _decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        batch_size, num_channels, num_frames, height, width = z.shape
+
+        if self.use_tiling and (width > self.tile_latent_min_width or height > self.tile_latent_min_height):
+            return self.tiled_decode(z, return_dict=return_dict)
+
+        frame_batch_size = self.num_latent_frames_batch_size
+        dec = []
+        for i in range(num_frames // frame_batch_size):
+            remaining_frames = num_frames % frame_batch_size
+            start_frame = frame_batch_size * i + (0 if i == 0 else remaining_frames)
+            end_frame = frame_batch_size * (i + 1) + remaining_frames
+            z_intermediate = z[:, :, start_frame:end_frame]
+            if self.post_quant_conv is not None:
+                z_intermediate = self.post_quant_conv(z_intermediate)
+            z_intermediate = self.decoder(z_intermediate)
+            dec.append(z_intermediate)
+
+        self._clear_fake_context_parallel_cache()
+        dec = torch.cat(dec, dim=2)
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
     @apply_forward_hook
-    def decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
+    def decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
         """
         Decode a batch of images.
 
@@ -1007,13 +1145,111 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             [`~models.vae.DecoderOutput`] or `tuple`:
                 If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
                 returned.
+        """
+        if self.use_slicing and z.shape[0] > 1:
+            decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)]
+            decoded = torch.cat(decoded_slices)
+        else:
+            decoded = self._decode(z).sample
 
+        if not return_dict:
+            return (decoded,)
+        return DecoderOutput(sample=decoded)
+
+    def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[3], b.shape[3], blend_extent)
+        for y in range(blend_extent):
+            b[:, :, :, y, :] = a[:, :, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, :, y, :] * (
+                y / blend_extent
+            )
+        return b
+
+    def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[4], b.shape[4], blend_extent)
+        for x in range(blend_extent):
+            b[:, :, :, :, x] = a[:, :, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, :, x] * (
+                x / blend_extent
+            )
+        return b
+
+    def tiled_decode(self, z: torch.Tensor, return_dict: bool = True) -> Union[DecoderOutput, torch.Tensor]:
+        r"""
+        Decode a batch of images using a tiled decoder.
+
+        Args:
+            z (`torch.Tensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
         """
-        if self.post_quant_conv is not None:
-            z = self.post_quant_conv(z)
-        dec = self.decoder(z)
+        # Rough memory assessment:
+        #   - In CogVideoX-2B, there are a total of 24 CausalConv3d layers.
+        #   - The biggest intermediate dimensions are: [1, 128, 9, 480, 720].
+        #   - Assume fp16 (2 bytes per value).
+        # Memory required: 1 * 128 * 9 * 480 * 720 * 24 * 2 / 1024**3 = 17.8 GB
+        #
+        # Memory assessment when using tiling:
+        #   - Assume everything as above but now HxW is 240x360 by tiling in half
+        # Memory required: 1 * 128 * 9 * 240 * 360 * 24 * 2 / 1024**3 = 4.5 GB
+
+        batch_size, num_channels, num_frames, height, width = z.shape
+
+        overlap_height = int(self.tile_latent_min_height * (1 - self.tile_overlap_factor_height))
+        overlap_width = int(self.tile_latent_min_width * (1 - self.tile_overlap_factor_width))
+        blend_extent_height = int(self.tile_sample_min_height * self.tile_overlap_factor_height)
+        blend_extent_width = int(self.tile_sample_min_width * self.tile_overlap_factor_width)
+        row_limit_height = self.tile_sample_min_height - blend_extent_height
+        row_limit_width = self.tile_sample_min_width - blend_extent_width
+        frame_batch_size = self.num_latent_frames_batch_size
+
+        # Split z into overlapping tiles and decode them separately.
+        # The tiles have an overlap to avoid seams between tiles.
+        rows = []
+        for i in range(0, height, overlap_height):
+            row = []
+            for j in range(0, width, overlap_width):
+                time = []
+                for k in range(num_frames // frame_batch_size):
+                    remaining_frames = num_frames % frame_batch_size
+                    start_frame = frame_batch_size * k + (0 if k == 0 else remaining_frames)
+                    end_frame = frame_batch_size * (k + 1) + remaining_frames
+                    tile = z[
+                        :,
+                        :,
+                        start_frame:end_frame,
+                        i : i + self.tile_latent_min_height,
+                        j : j + self.tile_latent_min_width,
+                    ]
+                    if self.post_quant_conv is not None:
+                        tile = self.post_quant_conv(tile)
+                    tile = self.decoder(tile)
+                    time.append(tile)
+                self._clear_fake_context_parallel_cache()
+                row.append(torch.cat(time, dim=2))
+            rows.append(row)
+
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                # blend the above tile and the left tile
+                # to the current tile and add the current tile to the result row
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent_height)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent_width)
+                result_row.append(tile[:, :, :, :row_limit_height, :row_limit_width])
+            result_rows.append(torch.cat(result_row, dim=4))
+
+        dec = torch.cat(result_rows, dim=3)
+
         if not return_dict:
             return (dec,)
+
         return DecoderOutput(sample=dec)
 
     def forward(

src/diffusers/models/transformers/cogvideox_transformer_3d.py

@@ -37,13 +37,20 @@ class CogVideoXBlock(nn.Module):
     Transformer block used in [CogVideoX](https://github.com/THUDM/CogVideo) model.
 
     Parameters:
-        dim (`int`): The number of channels in the input and output.
-        num_attention_heads (`int`): The number of heads to use for multi-head attention.
-        attention_head_dim (`int`): The number of channels in each head.
-        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
-        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
-        attention_bias (:
-            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        dim (`int`):
+            The number of channels in the input and output.
+        num_attention_heads (`int`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`):
+            The number of channels in each head.
+        time_embed_dim (`int`):
+            The number of channels in timestep embedding.
+        dropout (`float`, defaults to `0.0`):
+            The dropout probability to use.
+        activation_fn (`str`, defaults to `"gelu-approximate"`):
+            Activation function to be used in feed-forward.
+        attention_bias (`bool`, defaults to `False`):
+            Whether or not to use bias in attention projection layers.
         qk_norm (`bool`, defaults to `True`):
             Whether or not to use normalization after query and key projections in Attention.
         norm_elementwise_affine (`bool`, defaults to `True`):
@@ -147,36 +154,53 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
     A Transformer model for video-like data in [CogVideoX](https://github.com/THUDM/CogVideo).
 
     Parameters:
-        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
-        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
-        in_channels (`int`, *optional*):
+        num_attention_heads (`int`, defaults to `30`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, defaults to `64`):
+            The number of channels in each head.
+        in_channels (`int`, defaults to `16`):
             The number of channels in the input.
-        out_channels (`int`, *optional*):
+        out_channels (`int`, *optional*, defaults to `16`):
             The number of channels in the output.
-        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
-        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
-        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
-        attention_bias (`bool`, *optional*):
-            Configure if the `TransformerBlocks` attention should contain a bias parameter.
-        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
-            This is fixed during training since it is used to learn a number of position embeddings.
-        patch_size (`int`, *optional*):
+        flip_sin_to_cos (`bool`, defaults to `True`):
+            Whether to flip the sin to cos in the time embedding.
+        time_embed_dim (`int`, defaults to `512`):
+            Output dimension of timestep embeddings.
+        text_embed_dim (`int`, defaults to `4096`):
+            Input dimension of text embeddings from the text encoder.
+        num_layers (`int`, defaults to `30`):
+            The number of layers of Transformer blocks to use.
+        dropout (`float`, defaults to `0.0`):
+            The dropout probability to use.
+        attention_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in the attention projection layers.
+        sample_width (`int`, defaults to `90`):
+            The width of the input latents.
+        sample_height (`int`, defaults to `60`):
+            The height of the input latents.
+        sample_frames (`int`, defaults to `49`):
+            The number of frames in the input latents. Note that this parameter was incorrectly initialized to 49
+            instead of 13 because CogVideoX processed 13 latent frames at once in its default and recommended settings,
+            but cannot be changed to the correct value to ensure backwards compatibility. To create a transformer with
+            K latent frames, the correct value to pass here would be: ((K - 1) * temporal_compression_ratio + 1).
+        patch_size (`int`, defaults to `2`):
             The size of the patches to use in the patch embedding layer.
-        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
-        num_embeds_ada_norm ( `int`, *optional*):
-            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
-            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
-            added to the hidden states. During inference, you can denoise for up to but not more steps than
-            `num_embeds_ada_norm`.
-        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
-            The type of normalization to use. Options are `"layer_norm"` or `"ada_layer_norm"`.
-        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+        temporal_compression_ratio (`int`, defaults to `4`):
+            The compression ratio across the temporal dimension. See documentation for `sample_frames`.
+        max_text_seq_length (`int`, defaults to `226`):
+            The maximum sequence length of the input text embeddings.
+        activation_fn (`str`, defaults to `"gelu-approximate"`):
+            Activation function to use in feed-forward.
+        timestep_activation_fn (`str`, defaults to `"silu"`):
+            Activation function to use when generating the timestep embeddings.
+        norm_elementwise_affine (`bool`, defaults to `True`):
             Whether or not to use elementwise affine in normalization layers.
-        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon value to use in normalization layers.
-        caption_channels (`int`, *optional*):
-            The number of channels in the caption embeddings.
-        video_length (`int`, *optional*):
-            The number of frames in the video-like data.
+        norm_eps (`float`, defaults to `1e-5`):
+            The epsilon value to use in normalization layers.
+        spatial_interpolation_scale (`float`, defaults to `1.875`):
+            Scaling factor to apply in 3D positional embeddings across spatial dimensions.
+        temporal_interpolation_scale (`float`, defaults to `1.0`):
+            Scaling factor to apply in 3D positional embeddings across temporal dimensions.
     """
 
     _supports_gradient_checkpointing = True
@@ -186,7 +210,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
         self,
         num_attention_heads: int = 30,
         attention_head_dim: int = 64,
-        in_channels: Optional[int] = 16,
+        in_channels: int = 16,
         out_channels: Optional[int] = 16,
         flip_sin_to_cos: bool = True,
         freq_shift: int = 0,
@@ -304,7 +328,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
         encoder_hidden_states = hidden_states[:, : self.config.max_text_seq_length]
         hidden_states = hidden_states[:, self.config.max_text_seq_length :]
 
-        # 5. Transformer blocks
+        # 4. Transformer blocks
         for i, block in enumerate(self.transformer_blocks):
             if self.training and self.gradient_checkpointing:
 
@@ -331,11 +355,11 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin):
 
         hidden_states = self.norm_final(hidden_states)
 
-        # 6. Final block
+        # 5. Final block
         hidden_states = self.norm_out(hidden_states, temb=emb)
         hidden_states = self.proj_out(hidden_states)
 
-        # 7. Unpatchify
+        # 6. Unpatchify
         p = self.config.patch_size
         output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, channels, p, p)
         output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)

src/diffusers/pipelines/cogvideo/pipeline_cogvideox.py

@@ -332,20 +332,11 @@ class CogVideoXPipeline(DiffusionPipeline):
         latents = latents * self.scheduler.init_noise_sigma
         return latents
 
-    def decode_latents(self, latents: torch.Tensor, num_seconds: int):
+    def decode_latents(self, latents: torch.Tensor) -> torch.Tensor:
         latents = latents.permute(0, 2, 1, 3, 4)  # [batch_size, num_channels, num_frames, height, width]
         latents = 1 / self.vae.config.scaling_factor * latents
 
-        frames = []
-        for i in range(num_seconds):
-            start_frame, end_frame = (0, 3) if i == 0 else (2 * i + 1, 2 * i + 3)
-
-            current_frames = self.vae.decode(latents[:, :, start_frame:end_frame]).sample
-            frames.append(current_frames)
-
-        self.vae.clear_fake_context_parallel_cache()
-
-        frames = torch.cat(frames, dim=2)
+        frames = self.vae.decode(latents).sample
         return frames
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
@@ -438,8 +429,7 @@ class CogVideoXPipeline(DiffusionPipeline):
         negative_prompt: Optional[Union[str, List[str]]] = None,
         height: int = 480,
         width: int = 720,
-        num_frames: int = 48,
-        fps: int = 8,
+        num_frames: int = 49,
         num_inference_steps: int = 50,
         timesteps: Optional[List[int]] = None,
         guidance_scale: float = 6,
@@ -534,9 +524,10 @@ class CogVideoXPipeline(DiffusionPipeline):
             `tuple`. When returning a tuple, the first element is a list with the generated images.
         """
 
-        assert (
-            num_frames <= 48 and num_frames % fps == 0 and fps == 8
-        ), f"The number of frames must be divisible by {fps=} and less than 48 frames (for now). Other values are not supported in CogVideoX."
+        if num_frames > 49:
+            raise ValueError(
+                "The number of frames must be less than 49 for now due to static positional embeddings. This will be updated in the future to remove this limitation."
+            )
 
         if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
             callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
@@ -593,7 +584,6 @@ class CogVideoXPipeline(DiffusionPipeline):
 
         # 5. Prepare latents.
         latent_channels = self.transformer.config.in_channels
-        num_frames += 1
         latents = self.prepare_latents(
             batch_size * num_videos_per_prompt,
             latent_channels,
@@ -673,7 +663,7 @@ class CogVideoXPipeline(DiffusionPipeline):
                     progress_bar.update()
 
         if not output_type == "latent":
-            video = self.decode_latents(latents, num_frames // fps)
+            video = self.decode_latents(latents)
             video = self.video_processor.postprocess_video(video=video, output_type=output_type)
         else:
             video = latents

tests/models/transformers/test_models_transformer_cogvideox.py

+# coding=utf-8
+# Copyright 2024 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import torch
+
+from diffusers import CogVideoXTransformer3DModel
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    torch_device,
+)
+
+from ..test_modeling_common import ModelTesterMixin
+
+
+enable_full_determinism()
+
+
+class CogVideoXTransformerTests(ModelTesterMixin, unittest.TestCase):
+    model_class = CogVideoXTransformer3DModel
+    main_input_name = "hidden_states"
+
+    @property
+    def dummy_input(self):
+        batch_size = 2
+        num_channels = 4
+        num_frames = 1
+        height = 8
+        width = 8
+        embedding_dim = 8
+        sequence_length = 8
+
+        hidden_states = torch.randn((batch_size, num_frames, num_channels, height, width)).to(torch_device)
+        encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
+        timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device)
+
+        return {
+            "hidden_states": hidden_states,
+            "encoder_hidden_states": encoder_hidden_states,
+            "timestep": timestep,
+        }
+
+    @property
+    def input_shape(self):
+        return (1, 4, 8, 8)
+
+    @property
+    def output_shape(self):
+        return (1, 4, 8, 8)
+
+    def prepare_init_args_and_inputs_for_common(self):
+        init_dict = {
+            # Product of num_attention_heads * attention_head_dim must be divisible by 16 for 3D positional embeddings.
+            "num_attention_heads": 2,
+            "attention_head_dim": 8,
+            "in_channels": 4,
+            "out_channels": 4,
+            "time_embed_dim": 2,
+            "text_embed_dim": 8,
+            "num_layers": 1,
+            "sample_width": 8,
+            "sample_height": 8,
+            "sample_frames": 8,
+            "patch_size": 2,
+            "temporal_compression_ratio": 4,
+            "max_text_seq_length": 8,
+        }
+        inputs_dict = self.dummy_input
+        return init_dict, inputs_dict

tests/pipelines/cogvideox/test_cogvideox.py

@@ -125,11 +125,6 @@ class CogVideoXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             # Cannot reduce because convolution kernel becomes bigger than sample
             "height": 16,
             "width": 16,
-            # TODO(aryan): improve this
-            # Cannot make this lower due to assert condition in pipeline at the moment.
-            # The reason why 8 can't be used here is due to how context-parallel cache works where the first
-            # second of video is decoded from latent frames (0, 3) instead of [(0, 2), (2, 3)]. If 8 is used,
-            # the number of output frames that you get are 5.
             "num_frames": 8,
             "max_sequence_length": 16,
             "output_type": "pt",
@@ -148,8 +143,8 @@ class CogVideoXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         video = pipe(**inputs).frames
         generated_video = video[0]
 
-        self.assertEqual(generated_video.shape, (9, 3, 16, 16))
-        expected_video = torch.randn(9, 3, 16, 16)
+        self.assertEqual(generated_video.shape, (8, 3, 16, 16))
+        expected_video = torch.randn(8, 3, 16, 16)
         max_diff = np.abs(generated_video - expected_video).max()
         self.assertLessEqual(max_diff, 1e10)
 
@@ -250,6 +245,36 @@ class CogVideoXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
                 "Attention slicing should not affect the inference results",
             )
 
+    def test_vae_tiling(self, expected_diff_max: float = 0.2):
+        generator_device = "cpu"
+        components = self.get_dummy_components()
+
+        pipe = self.pipeline_class(**components)
+        pipe.to("cpu")
+        pipe.set_progress_bar_config(disable=None)
+
+        # Without tiling
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_without_tiling = pipe(**inputs)[0]
+
+        # With tiling
+        pipe.vae.enable_tiling(
+            tile_sample_min_height=96,
+            tile_sample_min_width=96,
+            tile_overlap_factor_height=1 / 12,
+            tile_overlap_factor_width=1 / 12,
+        )
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_with_tiling = pipe(**inputs)[0]
+
+        self.assertLess(
+            (to_np(output_without_tiling) - to_np(output_with_tiling)).max(),
+            expected_diff_max,
+            "VAE tiling should not affect the inference results",
+        )
+
 
 @slow
 @require_torch_gpu